[PATCH 2/2] staging: octeon: fix out-of-bounds reads in tx path

From: Ayush Mukkanwar

Date: Sat Sep 05 2026 - 11:37:08 EST


1. cvm_oct_xmit() and cvm_oct_xmit_pow() blindly trust
skb->protocol == htons(ETH_P_IP) and dereference the IP header
without verifying if the packet is long enough or if the header
is in the linear skb data area.
Fix by checking if skb_headlen is larger than
skb_network_offset + size of ip header.

2. cvm_oct_xmit_pow() copies packet data into a fixed size
hardware buffer using a hardcoded memcpy() size. If a packet
is smaller than it, memcpy() will read past the end of the
skb buffer. Fix this by using min_t() so that it does not read
past the end of the skb.

Fixes: 80ff0fd3ab64 ("Staging: Add octeon-ethernet driver files.")
Reported-by: Sashiko <sashiko-bot@xxxxxxxxxx>
Closes: https://sashiko.dev/#/message/20260615172734.42038-1-ayushmukkanwar%40gmail.com
Signed-off-by: Ayush Mukkanwar <ayushmukkanwar@xxxxxxxxx>
---
Note: This patch has only been compile tested. No runtime testing was
performed as I do not have access to Octeon hardware.

drivers/staging/octeon/ethernet-tx.c | 15 +++++++++++++--
1 file changed, 13 insertions(+), 2 deletions(-)

diff --git a/drivers/staging/octeon/ethernet-tx.c b/drivers/staging/octeon/ethernet-tx.c
index 5e536827f87a..452bc9fc8a5d 100644
--- a/drivers/staging/octeon/ethernet-tx.c
+++ b/drivers/staging/octeon/ethernet-tx.c
@@ -361,6 +361,8 @@ netdev_tx_t cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)

/* Check if we can use the hardware checksumming */
if ((skb->protocol == htons(ETH_P_IP)) &&
+ (skb_network_offset(skb) >= 0) &&
+ (skb_network_offset(skb) + sizeof(struct iphdr) <= skb_headlen(skb)) &&
(ip_hdr(skb)->version == 4) &&
(ip_hdr(skb)->ihl == 5) &&
((ip_hdr(skb)->frag_off == 0) ||
@@ -571,6 +573,14 @@ netdev_tx_t cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
work->packet_ptr.s.back = (copy_location - packet_buffer) >> 7;

if (skb->protocol == htons(ETH_P_IP)) {
+ if (unlikely(!pskb_may_pull(skb, ETH_HLEN + sizeof(struct iphdr)))) {
+ cvmx_fpa_free(packet_buffer, CVMX_FPA_PACKET_POOL, 0);
+ cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, 1);
+ dev->stats.tx_dropped++;
+ dev_kfree_skb_any(skb);
+ return NETDEV_TX_OK;
+ }
+
work->word2.s.ip_offset = 14;
work->word2.s.tcp_or_udp =
(ip_hdr(skb)->protocol == IPPROTO_TCP) ||
@@ -587,7 +597,7 @@ netdev_tx_t cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
* does.
*/
memcpy(work->packet_data, skb->data + 10,
- sizeof(work->packet_data));
+ min_t(unsigned int, skb->len - 10, sizeof(work->packet_data)));
} else {
work->word2.snoip.is_rarp = skb->protocol == htons(ETH_P_RARP);
work->word2.snoip.is_arp = skb->protocol == htons(ETH_P_ARP);
@@ -596,7 +606,8 @@ netdev_tx_t cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev)
work->word2.snoip.is_mcast =
(skb->pkt_type == PACKET_MULTICAST);
work->word2.snoip.not_IP = 1; /* IP was done up above */
- memcpy(work->packet_data, skb->data, sizeof(work->packet_data));
+ memcpy(work->packet_data, skb->data,
+ min_t(unsigned int, skb->len, sizeof(work->packet_data)));
}

/* Submit the packet to the POW */
--
2.54.0